BrowserOS neo: a second Chromium browser that MCP agents drive with your real logins

A second browser that agents use while Chrome stays your main one

BrowserOS neo is explicit that it is not a Chrome replacement. It is a secondary browser that sits next to Chrome and is built to be friendly to agents. The idea is simple: you keep browsing in Chrome, and your AI agents get their own browser with their own tabs, where you can watch them work or replay what they did afterwards.

The repository is browseros-ai/BrowserOS, written mainly in TypeScript at the agent layer and licensed under AGPL-3.0. The project calls itself free, open source, and running everything on your machine. Its own description positions it as an open-source alternative to ChatGPT Atlas, Perplexity Comet and Dia.

The consequence of the second-browser design is that you run two Chromium-based browsers side by side. That is the point: agent activity stays out of the tabs you are working in. But it also means BrowserOS neo is not where you should expect your day-to-day browsing to happen, and comparing it feature for feature with Chrome misses what it is for.

One-click Chrome import hands agents your real accounts

The setup flow has four steps: install BrowserOS neo, import from Chrome, connect your agent, and give it a task. The import step is the one that makes the product work and the one that deserves the most thought. One click brings over your logins, bookmarks and extensions, and because BrowserOS neo is a Chromium fork, Chrome extensions run and your passwords and settings come across too.

After that, agents work with your real accounts from the first task. The examples the project gives are all logged-in work: posting to LinkedIn or Twitter/X, queuing posts and pulling engagement numbers, clearing an inbox and unsubscribing from junk email, updating a CRM, filing expenses, and pulling reports from internal tools.

This is also the main risk. An agent acting in a logged-in browser can post, delete and submit as you. BrowserOS neo gives you visibility through the live dashboard and replays, but those show you what happened; they do not stop an agent from doing it. If an account carries real consequences, such as a company social profile or an expense system, think about whether it belongs in the imported profile at all.

Tasks start in the agent over MCP, not in the browser

BrowserOS neo does not ship its own assistant. Tasks are given from your agent, not from the browser: you type something like "Book me the cheapest flight to London" into the agent, and the work appears in a new tab in BrowserOS neo.

Connection goes through MCP. Any AI that speaks MCP can use it, and the browser detects Claude Code, Codex, Cursor, VS Code, OpenClaw and Hermes on your machine for one-click connection. The FAQ adds Zed, OpenCode and Antigravity to the one-click list, and the comparison section names Cowork among the agents it works with. The project says it built tools optimised for web use and claims that, for the same task, BrowserOS neo uses fewer tokens than Claude's Chrome extension or the Codex browser; treat that as the vendor's claim until you measure it on your own tasks.

Parallelism is built in. You can fire off several tasks at once, each agent works in its own tab, and you keep browsing meanwhile. The new tab page acts as a live dashboard showing which site each agent is on, what it is doing and how far along it is. The MCP setup is documented at docs.browseros.com/neo/mcp.

Every session is recorded as video under ~/.browserclaw/

Each agent session is saved to disk as a scrubbable video with a step-by-step action timeline, so you can rewind and see exactly what happened. Sessions, screenshots, history and settings all live under `~/.browserclaw/` and, according to the project, never upload. The replay feature is documented at docs.browseros.com/neo/audit-and-replay.

This is the strongest safety feature in the product, because it turns an opaque agent run into something you can audit after the fact. It also has costs. Recordings of logged-in sessions capture whatever the agent saw, including inbox contents, CRM records and internal reports. Anyone who can read `~/.browserclaw/` on that machine can see them, so the directory deserves the same care as the accounts themselves. Video recordings of every session also accumulate on disk, so check how the privacy and replay docs handle retention before running agents all day.

Separately from the local data, BrowserOS neo sends anonymous product-usage events: agent connect and disconnect, version, and OS. The project states that it never sends URLs, page content, prompts, tool results or screenshots, and the events can be switched off with one toggle in Settings.

Running on 127.0.0.1 instead of a datacenter or a headless subprocess

BrowserOS neo defines itself against three categories of alternatives, and each comparison explains a design choice.

Against headless drivers: Playwright and agent-browser spin up a fresh Chrome subprocess with no logins. That is fine for CI and useless for work that needs your logged-in state, such as reading your inbox. BrowserOS neo imports logins once and persists them across sessions.

Against cloud browsers: browser-use and browserbase run in a datacenter, so logging in is a pain, and sites like Twitter and LinkedIn block traffic from datacenter IPs. BrowserOS neo runs on your machine, on `127.0.0.1`.

Against locked-in AI browsers: Atlas, Comet and Dia only work with their own AI. BrowserOS neo works with the agents you already use and pay for.

The flip side is that local execution ties agents to your machine. Tasks run where BrowserOS neo is installed, so if that machine is asleep or offline, nothing runs. For headless CI jobs, the project's own framing says Playwright is the better tool.

Desktop only: macOS, Windows and Linux x64

BrowserOS neo runs on macOS, Windows and Linux, with Linux builds limited to x64, and system requirements match Google Chrome. The recommended macOS install goes through a Homebrew tap:

brew tap browseros-ai/tap
brew install --cask browseros-neo

Direct downloads are available from cdn.browseros.com: BrowserOS_neo.dmg for macOS, BrowserOS_neo_installer.exe for Windows, and BrowserOS_neo.AppImage or BrowserOS_neo.deb for Linux x64.

That platform list has no Android, iOS or other mobile build, so searches for BrowserOS on Android or BrowserOS mobile will not find a version in this repository. Linux users on ARM hardware are also outside the supported set, since only x64 builds are offered.

The four-step setup goes straight from installing to importing your Chrome profile, so the first decision after install is which logins you bring across. Enterprise deployment is handled by email at [email protected] rather than through a separate edition in the repository.

A Chromium fork and a TypeScript and Rust agent platform in one monorepo

BrowserOS neo ships from a single monorepo with two main subsystems. The browser is a Chromium fork written in C++ and Python, under packages/browseros/. The agent platform is TypeScript and Rust, which is why the root carries a .cargo/ directory next to packages/, tools/, skills/ and updates/. A LICENSE.ungoogled_chromium file sits beside the main AGPL-3.0 LICENSE, and the repository uses git submodules, per its .gitmodules file.

Releases are cut per component rather than as one version. Recent tags include BrowserClaw Server v0.0.89 and v0.0.90 on 28 and 30 October 2026, and the BrowserOS neo Extension v0.2.54.0 on 29 October 2026. The last push was on 2 November 2026, so the project is under active development.

For contributors, AGPL-3.0 matters: modified versions offered to users over a network carry source obligations. The root also holds CLA.md and a signatures/ directory, so read CLA.md and CONTRIBUTING.md before opening a pull request. Bugs go to the GitHub issue tracker, and support runs through Discord and Slack.

Editorial conclusion

BrowserOS neo suits people who already pay for Claude Code, Codex, Cursor or another MCP agent and want it to work inside logged-in sites such as LinkedIn, a CRM or an inbox, on a desktop running macOS, Windows or Linux x64. It does not suit anyone looking for a Chrome replacement, a mobile or Android browser, or a CI automation driver, where Playwright fits better. Before importing your Chrome profile, decide which accounts you are willing to let an agent act on, read the privacy policy at docs.browseros.com/neo/privacy, and turn off the usage events in Settings if you do not want them sent.

Frequently asked questions

Official sources

1. browseros-ai/BrowserOS on GitHub
2. License: AGPL-3.0
3. Project website
4. README
5. Releases